f-Domain-Adversarial Learning: Theory and Algorithms for Unsupervised Domain Adaptation with Neural Networks

David Acuna, Guojun Zhang, Marc T Law, Sanja Fidler

TL;DR

This paper introduces "f-Domain-Adversarial Learning" (f-DAL), a new framework for unsupervised domain adaptation that unifies existing theories and algorithms under the umbrella of f-divergences. By deriving new generalization bounds and framing the optimization as a Stackelberg game, f-DAL achieves state-of-the-art results on challenging benchmarks, demonstrating the power of a broader theoretical perspective.

What Problem Does This Paper Address?

Unsupervised domain adaptation (UDA) is crucial when a model trained on one data distribution (source domain) needs to perform well on a different but related distribution (target domain) without access to labeled target data. Existing theoretical bounds for UDA often rely on divergence measures that are difficult to optimize in practice, creating a gap between theory and the methods used in state-of-the-art deep learning approaches. This disconnect hinders progress and limits the understanding of why certain methods work. The paper aims to bridge this gap by developing a more general theoretical framework that directly informs practical, high-performing algorithms.

Methods

The core of this work is the development of f-Domain-Adversarial Learning (f-DAL). Key methodological contributions include:

  • Unified Theoretical Framework: The paper derives new generalization bounds for UDA based on a variational characterization of f-divergences. This approach generalizes previous theoretical results (like those based on HΔH divergence) and accommodates a wide range of divergence measures used in practice.
  • General Algorithm for f-Divergences: They propose a general algorithmic framework that can leverage any f-divergence for domain-adversarial learning. This allows for exploring a broader set of divergence measures than previously considered.
  • Stackelberg Game Formulation: The optimization problem in f-DAL is recognized as a Stackelberg game. This insight allows the authors to leverage advanced game optimization techniques, specifically Aggressive Extra-Gradient (AExG), which are shown to improve convergence speed and stability.
  • Per-Category Domain Classifier: Unlike some prior methods that use a single domain classifier, f-DAL implicitly uses a per-category domain classifier, which is shown to be more effective.

Datasets

The paper evaluates its framework on two widely used benchmark datasets for domain adaptation:

  • Office-31: Contains 4,652 images from three domains (Amazon, Webcam, DSLR) across 31 categories.
  • Office-Home: A more challenging dataset with 15,500 images from four domains (Artistic images, Clip Art, Product images, Real-world images).

No new datasets were introduced.

Key Insights and Ideas

  • The Power of f-Divergences: The paper demonstrates that the entire family of f-divergences can be effectively used for domain adaptation, not just specific ones like Jensen-Shannon (JS) or HΔH. This opens up new avenues for research and algorithm design.
  • Pearson χ² as a Strong Performer: Surprisingly, the Pearson χ² divergence, which was not previously explored in domain-adversarial learning, achieves state-of-the-art results. This highlights the benefit of exploring less conventional divergences.
  • Stackelberg Equilibrium for Optimization: Recognizing the optimization as a Stackelberg game is a crucial theoretical insight. It justifies the use of advanced optimizers like AExG, leading to practical performance gains and faster convergence.
  • Theory-Algorithm Alignment: By deriving a more comprehensive theoretical framework, the paper successfully bridges the gap between theory and practice, leading to algorithms that are both theoretically grounded and empirically superior.

Key Results

  • State-of-the-Art Performance: f-DAL, particularly when using the Pearson χ² divergence and the AExG optimizer, achieves state-of-the-art accuracy on both the Office-31 and Office-Home datasets.
  • Improved Performance with AExG: The Aggressive Extra-Gradient (AExG) optimizer, derived from the Stackelberg game formulation, consistently shows performance gains and faster convergence compared to standard optimizers like GDA, especially for certain divergences.
  • No Extra Hyperparameters: Impressively, f-DAL achieves its SOTA results without requiring additional hyperparameters beyond those typically used in standard domain adaptation methods.
  • Generalization Across Divergences: The framework demonstrates that various f-divergences can be used effectively, with Pearson χ² showing particular promise.

Limitations and Future Work

  • Hyperparameter Tuning for Specific Divergences: While the framework itself doesn't require extra hyperparameters, some divergences (like γ-weighted JS) might require additional tuning of γ, which the authors chose not to pursue in their experiments for simplicity.
  • Further Exploration of Divergences: The paper suggests that further investigation into other f-divergences and their properties could yield even better results.
  • Advanced Optimization Techniques: While AExG shows promise, exploring other advanced game optimization techniques could further enhance performance.
  • Theoretical Extensions: The paper mentions that the framework could be extended to incorporate additional techniques like CDAN or MDD, which are currently used as separate methods.

Summary updated

Abstract

The problem of unsupervised domain adaptation arises in a variety of practical applications where the distribution of the training samples differs from those used at test time. The existing theory of domain adaptation derived generalization bounds based on divergence measures that are hard to optimize in practice. This has led to a large disconnect between theory and state-of-the-art methods. In this paper, we propose a novel domain-adversarial framework that introduces new theory for domain adaptation and leads to practical learning algorithms with neural networks. In particular, we derive a novel generalization bound that utilizes a new measure of discrepancy between distributions based on a variational characterization of f-divergences. We show that our bound recovers the theoretical results from Ben-David et al. (2010a) as a special case with a particular choice of divergence, and also supports divergences typically used in practice. We derive a general algorithm for domain-adversarial learning for the complete family of f-divergences. We provide empirical results for several f-divergences and show that some, not considered previously in domain-adversarial learning, achieve state-of-the-art results in practice. We provide empirical insights into how choosing a particular divergence affects the transfer performance on real-world datasets. By further recognizing the optimization problem as a Stackelberg game, we utilize the latest optimizers from the game optimization literature, achieving additional performance boosts in our training algorithm. We show that our f-domain adversarial framework achieves state-of-the-art results on the challenging Office-31 and Office-Home datasets without extra hyperparameters.
BibTeX
@misc{acuna2020f,
  author = {David Acuna and Guojun Zhang and Marc T Law and Sanja Fidler},
  title = {{f-Domain-Adversarial Learning: Theory and Algorithms for Unsupervised Domain Adaptation with Neural Networks}},
  year = {2020},
  url = {https://openreview.net/forum?id=WqXAKcwfZtI},
  howpublished = {OpenReview},
}
Figure